The Best Fitness App I’ve Ever Used Is a Markdown File

Early on, I photographed my lunch, typed a few words, and hit send. Claude estimated the nutrition, filed it as a row in a table, and then did something I hadn’t asked for: it connected the protein on the plate to a pattern we’d already discussed — my tendency to get hungry late in the evening — and told me this should help.

That was the first sign I was on to something. By my third Saturday review, the system itself had changed: my process adjusted, training rules updated, nuances about my eating habits added.

I had written no code. I was using an AI agent, an operating manual in plain English, and a handful of files — and for the first time in my life, I’d stuck with a fitness tracker long enough to improve it. There is no fitness app anywhere in the loop. Just Claude, with access to CSV and markdown files in the cloud.

What it does for me is simple to state:

1. I log every meal, and it keeps me on my daily protein and calorie targets, adjusting the advice as the day goes: what dinner should look like given what lunch was.
2. I log every workout, and it tells me when to add weight, following progression rules we wrote together, and what to watch for given my current and past injuries.
3. It reviews all of it, in the moment and again every Saturday, against the body-recomposition plan the whole system is built around.

Every app lost me in a session or two

I’ve abandoned every fitness tracker I’ve ever tried, and not for lack of interest. I’m exactly who these apps are built for: a former competitive athlete, back to serious training in middle age, with a real appetite for data on training, nutrition, and recovery. I wanted to be tracked.

The failure was always the same. Every app demanded organizing work from me, every time: searching a food database mid-meal, tapping through exercise pickers between sets, squeezing what I’d done into form fields built by someone who was not me. If I wanted to track something the app didn’t anticipate — too bad. And after all that entry, the analysis was still my job.

Describing a meal or a workout is easy. The pain was forcing that description into someone else’s structure, then doing the thinking myself. AI let me invert that: I enter information however it works for me, and the machine does the structuring and the analysis.

What I built: the manual is the software

Over two conversations on two evenings — three hours at the absolute most — I set up a personal tracking system in Claude Cowork, the side of Claude that runs agent tasks against a shared project. Claude was both my design partner and the agent that would operate it, with a head start: months of periodic conversations about my training — exercises, scheduling, old injuries, goals.

Today the system is one project. This is all of it:

Three kinds of files — rules, context, data — a general model, and me. The model brings general knowledge; everything specific to my training lives in prose I can read, question, and change. Nothing in the architecture is specific to fitness, which is the part I’d underline if you build products for a living. I’ll come back to it.

My routine: one agent thread per week, started at my desktop and fed from my phone. Claude files each entry, flags anything worth discussing, and answers a question from the gym floor with the week already in view. On Saturday, a phone alarm reminds me to ask for the review.

The build ran on simple tickets: numbered steps, done-criteria (“when finished, list every file so I can verify”), and a closing fence — do not change anything else. I even had Claude write the tickets. The only skill the whole thing required was a precise idea of what I wanted.

The product work was judgment, not building

Two evenings built the loop. The product work happened in the three weeks after, and almost none of it was building. When any change costs a sentence in a manual, the whole job is judgment: what to measure, what to tolerate, and who gets to change the rules.

Decide what accuracy is for. Photo-based nutrition estimates are easy to enter and noticeably imprecise. Rather than chase precision, the manual states what the numbers are for — trend direction and protein consistency — and rules everything else out of scope. Missing entries are gaps, never zeros. A tolerance you’ve written down is a design decision; one you haven’t is an argument waiting to happen.

Measure the behavior you actually want. A twice-weekly shoulder routine had picked up a progression rule I never asked for; the ambition was the metric’s, not mine. We removed it, and the metric became whether the two sessions happened. Choosing what deserves measuring is the product work, even when the change takes a sentence.

Let the rules bend, but guard the goalposts. Categories are editable — the system flagged a planned, bounded dessert as a late-night eating incident and I overruled it — but editable categories are also a machine for excusing anything. So two guardrails: my account goes in the record next to the evidence that might complicate it, and no metric changes definition without my agreement. Claude can propose. It can’t quietly move the goalposts.

Change one thing at a time. When three lifts were ready for heavier weights, the review staggered them so any discomfort the next morning would be interpretable. The log shows exactly that. It’s the payoff I always wanted from tracking: a decision, without a second session to produce it myself.

Three Saturday reviews in, I’ve stuck with a tracker longer than ever before. I have no need for a fitness app again. More precisely: the bar for what I’d even use for free — forget pay for — is now sky high.

I am still the QA department

The flexibility has a price: I have to check what comes back. None of it was frictionless — sandboxed tasks that forget files unless told to save them, a phone app that couldn’t start a thread, scheduling that didn’t work — but every fix lives in plain English: a rule in the manual, a habit in the routine. The errors that matter come in three kinds.

Estimates fail silently. A dinner ran about 450 calories high because miniature samosas were counted as full-sized ones. The fix was a serving-size rule in the manual — but the number looked fine until I questioned it. Plausible output is not verified output.

Explanations outrun the evidence. One review attributed a weight change to water on the strength of estimated intake, which can’t settle that question; another named one day the week’s only protein shortfall after listing two. A persuasive paragraph can coexist with a contradiction a few paragraphs away. Confidence is a writing style, not a signal.

Editable prose drifts. I asked for short reviews; they grew. A revised calorie policy landed in one file while the old target survived in two others. When the rules live in text, repetition gives old rules places to hide, and someone has to own the maintenance.

Two evenings got the first version running. They didn’t buy a finished product or eliminate upkeep. Sample size is one, and I’m not claiming a change in body composition. What I can claim is sustained use, feedback I act on, and a system I’ve been willing to maintain — as someone who enjoys writing an operating manual, which most people won’t.

The expertise layer is a shrinking moat

The obvious objection: fitness apps are adding AI too. Won’t they catch up?

I don’t think so, and the reason is the most resonant thing I’ve read about AI. Rich Sutton’s “The Bitter Lesson” argues that in AI research, general methods that scale with computation have repeatedly beaten approaches built on hand-encoded human expertise. Researchers keep adding what they know; the general method keeps winning anyway.

Sutton was writing about research, but it applies to product. Fitness apps will take a general model, wrap it in the app’s schema, and layer expertise on top: a curated food database, a coach’s programming logic, a trainer’s rules of thumb. That layer is the pitch — and it’s exactly the hand-encoded knowledge Sutton is describing.

In my case it wasn’t needed. A general model with no food database, no coaching logic, and no trainer’s rules co-designed the whole system with me: the categories, the progression rules, the review format, the guardrails. Whatever expertise it took, the model already had. The only thing that made the system fit my life was the manual I wrote — which the app can’t have.

So the real comparison isn’t app versus no app. It’s a general model reading the user’s own context, against the same model reading someone else’s expertise through a fixed schema.

The practical version, for anyone deciding what to build: the expertise layer is a shrinking moat. The model belongs to someone else and the context belongs to the user. What a product can still own is the specification — sensible defaults, taxonomies that hold up, the judgment about what to measure — and the plumbing my setup lacks: reliable records, connected devices, one place to change a rule. A smaller thing than a decade of features, and the part a general model can’t supply on its own.

The part that should worry software people

Look at Figure 1 one more time and swap the nouns: the manual becomes a sales playbook, the context a territory, the tables contacts and deals. Which reopens the question that started me down this path: what is most software, really, but a graphical interface for viewing and manipulating some data? A CRM, at bottom, exists so employees can complete tasks against company data. When posing the task to an agent returns the completed task, what exactly is the interface for?

A personal tracker proves nothing about enterprise software, which has moats mine doesn’t — multi-user state, integrations, compliance, the political weight of being the system of record. I’m not writing that obituary.

But the mechanism that killed my demand doesn’t respect category boundaries; it pressures any product that is mostly a rigid schema standing between one person and their data. Single-user consumer software is first in line, and note who defects first: motivated power users, the segment these products monetize best. I’m an early adopter — this took fluency with frontier tooling most people don’t have yet. It also took me two evenings, and that bar drops every month.

Once a regular person can do what I did — and we’re already close — a whole bunch of the companies selling it are going out of business.